V3 turns the agent into a twin: it remembers you permanently, disagrees with dated evidence, challenges you when you drift from your own goals — and earns, approval by approval, the right to act in your name. You grow. It grows. Both, measurably.
Every other agent resets to zero each session. A twin compounds — and it compounds you with it.
🧑How you grow
It keeps your promises honest — stated goals get compared to your actual behavior, weekly, with evidence
It confronts your drift instead of flattering it: "40 days, no movement — three moves, or tell me it stopped mattering"
It holds the questions you'd rather forget as open threads — resolved deliberately, never dropped
It answers "why do you think that about me?" with your own words, dated
🤖How it grows
Every correction becomes permanent memory — a corrected mistake never repeats
Every approval extends a trust streak; five clean approvals and it earns an action class outright
Every back-off is obeyed and remembered — it learns what NOT to raise, with your reasons
Positions it holds about you carry evidence and revision history — it can be wrong, revise out loud, and be audited
% of delegated work completed without a human touch · autonomy is earned, never granted — one edit revokes it
What V3 Adds
Mechanisms that carry the relation.
Not a persona file that sounds consistent and knows nothing. Structure — because for a twin, care is structure or it's absent.
New
📜
The Founding Charter
CHARTER.md — a constitution you own, in your workspace. You grant which domains it may confront you about, evidence thresholds, and back-off terms. Deny by default; amended by ceremony; hot-reloaded on every action.
New
🧠
Constitutional Memory
The founding episode, turning points, accepted corrections, granted authorities — permanent, never summarized away, present in every prompt. Carried state always arrives with its date: no date, nothing carried.
New
🧭
Positions & Dissent
It holds evidence-backed claims about you and pushes back once when you contradict yourself — citing dates. Overruled? It complies and keeps the question open. Corrected? Revised out loud, prior claim kept in the audit trail.
New
⚖️
Charter-Gated Confrontation
At most one challenge a week, only through a deterministic gate: granted domain, evidence threshold met. Back off and it obeys AND remembers; the third back-off retires the goal forever. Timeout records nothing — silence is never consent.
New
🎓
Earned Delegation
Clean-approval streaks graduate action classes to act-without-asking — only within charter scope, only when you're unreachable (by default), always receipted. One edit revokes everything and the streak restarts from zero.
New
🧾
The Morning Digest
Everything done in your absence, each action with a receipt, delivered at 09:00. Nothing happened? No message. A twin is a partner, not a notification app.
New
🖥️
The Twin Home
One view in the app: what needs you (answer with the same buttons as Telegram), the relationship's full activity record, open threads, positions with evidence, live trust streaks, and your charter.
Teams
🛡️
Conduct, Never Content
In team deployments every member gets their own isolated twin — and admins see what twins did (actions, receipts, freeze switches), never what they know. Memories, reasons, and goals stay owner-private, structurally.
The Difference
Progress your business can actually see.
You choose the accountability domains in a charter you own. The twin shows — per domain — how well it understands the pattern, how close it is to being cleared to act, whether its nudges land, and which way each trends. Board-ready, not a black box.
You set what it's allowed to hold people to — a founder picks pipeline & runway; a delivery lead picks plan & client commitments — and get an auditable read of whether the AI is genuinely learning the operation: positions revised, corrections that stick, autonomy earned and sustained, nudges that land.
Evolution
Assistant → Workforce → Twin
V3 ships as the twin-connect update inside both editions — V1 for yourself, V2 for your whole team.
Capability
V1 — Agent
V2 — Team Workforce
V3 — Digital Twin
Memory
Vector + RAG
5-layer (episodic, biographic, counterparty)
✓ + Constitutional — permanent, dated, never compressed
🧬 V3 twin update — rolling out to V1 & V2🔒 Team edition invite-only
Get first access to the twin update, onboarding for your charter, and direct founder contact.
Your twin only ever answers to you — starting with this list.
Open source · Self-hosted · Free to run
The agent that works for you.
AI Partner V1 is a self-hosted, open-source autonomous agent. Give it a goal in plain English — it researches, codes, generates documents, and delivers the finished result on your own machine. Want the twin that grows with you?Meet V3 →
Agent pauses before running any script. Approve or reject via Telegram inline buttons. Also handles CAPTCHA hand-off, clarifications, and mid-execution input requests.
🌐
Live Browser Automation
Full browser with 15fps live screencast. SPA-aware rendering, stealth mode. CAPTCHA detected? Agent pauses, you solve in the live panel, it resumes.
🐳
Sandboxed Docker Execution
Every script runs inside a persistent container with Python, Node.js, and Bash. pandas, yfinance, matplotlib pre-loaded. Network isolation per-goal.
📄
Document Generation
PDF via Chromium, multi-sheet XLSX, PowerPoint, and DOCX — all production-ready, tracked in the dashboard, and instantly downloadable.
⚡
Skill Learning
After a successful goal, the system generalizes the solution into a reusable parameterized template. Versioned, deduplicated, and promotable to a callable tool.
🤖
Multi-Agent Delegation
Orchestrators spin up to 5 parallel sub-agents with budget controls. When a research agent exhausts its cap, it writes a handoff summary — the parent picks up seamlessly.
🕑
Scheduler + Proactive Agenda
Cron-expression tasks persist across restarts. The Proactive Agenda reads your task list and memory — uses LLM judgment to pick the best action per tick.
🧬
Vector Memory + RAG
Episodic log, vector search, persona traits, and periodic consolidation. Works with Ollama, OpenAI, Cohere, or pure-JS TF-IDF fallback — zero API keys required.
🔌
17 Built-in MCP Servers
Browser, Shell, GitHub, Gmail, Calendar, Drive, Notion, Trello, Twitter, Spotify, Apify, Image Gen, Knowledge Base, Scheduler, and more. Plus any external MCP server.
How It Works
Goal in. Result out.
A 5-stage pipeline that turns natural language into validated, delivered outcomes.
1
Goal Extraction
Your natural-language goal is converted into a structured definition with typed, measurable success criteria — file existence, content patterns, delivery receipts — before a single action is taken.
2
Task Decomposition
Complex goals are broken into ordered sub-tasks with declared outputs. The right specialist agent is auto-selected by keyword or explicitly @mentioned.
3
Parallel Execution
Agents reason, pick tools, execute, and assess in a tight loop. Up to 5 sub-agents run in parallel. When a script fails, a semantic repair engine fixes it and retries.
4
Criteria-Based Validation
After every iteration the system checks real outcomes. The agent cannot self-report "done" — it's done when filesystem, content, and messaging evidence all pass.
5
Delivery + Checkpoint
Results delivered to Telegram, Discord, Slack, or your workspace. State is checkpointed — a restart picks up exactly where it stopped, all artifacts intact.
Agent Profiles
16 specialists. Zero generic.
@web-researcher
Web Researcher
Research
@fact-checker
Fact Checker
Research
@trend-spotter
Trend Spotter
Research
@python-dev
Python Developer
Dev
@node-dev
Node.js Developer
Dev
@debugger
Debugger
Dev
@shell-op
Shell Operator
Dev
@fin-analyst
Financial Analyst
Data
@data-analyst
Data Analyst
Data
@excel-builder
Excel Builder
Data
@report-gen
Report Generator
Content
@summarizer
Summarizer
Content
@tech-writer
Tech Writer
Content
@prompt-arch
Prompt Architect
Content
@task-planner
Task Planner
Content
@telegram-rep
Telegram Reporter
Delivery
Full Capabilities
What it can actually do
🎯
Criteria-Based Completion
Goals close only when real-world evidence passes — no self-reporting allowed.
🔁
Dynamic Replanning
Fails a step? Replans, tries alternatives, semantically repairs scripts, retries.
💾
Checkpoint & Resume
State saved after each iteration. Restarts pick up exactly where they stopped.
📲
Telegram as Control Panel
Send goals, approve scripts, receive files — all from Telegram, no UI needed.
Points to local Ollama by default. Start free, add keys progressively.
Integrations
Connected to everything
One env var to activate any integration. Auto-discovered at startup.
💬
Telegram
Tasks, approvals, files
🎮
Discord
Text & file delivery
💼
Slack
Socket Mode
📱
WhatsApp
QR auth
🐙
GitHub
Issues, PRs, code
📝
Notion
Read, create, query
📅
Google Calendar
Events + triggers
📧
Gmail
Send, search, trigger
💾
Google Drive
Search, create
🐦
Twitter / X
Search, post, reply
🎵
Spotify
Play, pause, queue
🕷️
Apify
Proxy scraping
🎨
Image Gen
DALL-E 3 / Stability
Demo Flows
Real tasks. Real outcomes.
1
Daily Tech Briefing → Telegram
"Research today's top AI stories from HN + GitHub. Write a report and send to Telegram."
Full pipeline: research → PDF generation → validated Telegram delivery. Goal closes only after the file actually arrives.
Trend SpotterWeb SearchReport GenPDFTelegram
2
Live Stock Portfolio Snapshot
"Fetch closing prices for RELIANCE.NS, TCS.NS, INFY.NS. Build Excel with price, change%, 5-day chart."
Python + yfinance runs in Docker sandbox. Styled multi-sheet Excel with embedded chart. File criterion confirms output.
Financial AnalystyfinanceExcel BuilderXLSX
3
CAPTCHA-Protected Research
"Go to LinkedIn and find the top 5 posts about AI agents this week."
Browser navigates → CAPTCHA detected → Inspector opens → you solve in live preview → agent resumes and extracts content.
BrowserCAPTCHA HITLLive Screencast
4
Automated Weekly GitHub Report
Every Monday 9am → "Summarize commits, note open PRs, post summary to Slack."
Scheduler fires → GitHub data fetched → report written → Slack delivery. Set once, runs forever.
SchedulerGitHubReport GenSlack
5
Skill Learning: Instant Second Run
"Fetch NIFTY50 data and plot the last 30 days."
First run: Python from scratch. After success, system extracts a reusable template. Second run: template fills in → instant execution.
Skill LearningTemplatesTool Promotion
🚀 The Team Edition
🔒 Private — not open-source
📅 Private beta — invites rolling out
An AI workforce for your whole team.
V2 is the multi-tenant, governed edition. Deploy once for your company — every member gets isolated agents, admins get full audit and control, and an opt-in AI proxy can act on your behalf, always behind your approval. Invite-only.
No scripts. No staging. AI Partner joins a live call, listens in real time, and responds.
Who Is This For?
Built for busy professionals
🎯 Entrepreneurs & Founders
Double-booked by default? AI Partner attends meetings you can't make, follows up on your behalf, monitors your market while you sleep. Your time is your revenue — stop wasting it on tasks that don't need you.
💼 Corporate CXO & Senior Leaders
Spend 6-8 hours daily in meetings most don't need your judgment — just your presence. Send the proxy. It attends, flags blockers against your OKRs, captures decisions, and reports back in minutes.
What V2 Adds
Everything V1 does. Plus this.
New
🎭
Inbox & Chat Proxy — With Your Approval
AI Partner drafts replies to your emails, Slack DMs, and Telegram messages in your style. Nothing sends on its own — every action on your behalf is gated by AuthorityPolicy and a one-tap approval, and AI involvement is always disclosed.
New
📹
Meeting Assistant — In the Room For You
Joins your Meet/Teams/Zoom calls with AI presence disclosed, listens, takes notes, and contributes only what you've authorized — then generates summaries and extracts action items automatically.
New
📡
Proactive Heartbeat Engine
V2 doesn't wait to be asked. A background loop evaluates what to do every 15-30 minutes using your goals, memory, and real-time context.
Enhanced
🔧
100+ Tools, 36 MCP Servers
From 17 servers in V1 to 36 in V2. Plus 4-tier computer use (DOM → Vision → Container Desktop → Host Control) with live CAPTCHA handoff.
New
📱
Phone Assistant — Voicemail + Outbound
Screens calls for you, then records, transcribes, and summarizes. Places approved outbound calls from your Twilio number — every call authorized by you, with AI disclosure on connect.
Enhanced
🧠
5-Layer Memory System
Episodic + Biographic + Counterparty + Vector + RAG. Understands who's who, what you've discussed, and maintains relationship context across all channels.
Enhanced
🤖
18+ LLM Providers
From 8 providers in V1. Now including NVIDIA NIM, Cerebras (~100k tokens/s), OpenRouter (100+ models), MiniMax, and LiteLLM Proxy.
New
🔐
Authority Policy + HITL
Declarative rules define what the agent does automatically vs. what needs your approval. Telegram inline buttons for every gray-zone decision.
Built for Teams
AI you can actually roll out.
V2 is multi-tenant and governed from the core — the reason it's safe to deploy across a whole company.
👥
Multi-Tenant by Design
One install hosts your whole team. Every user's chats, files, memory, and screenshots are fully isolated — no cross-user leak, enforced at the data, event, and identity layers.
🛡️
Admin Console
A control room with a live activity feed, users & invites, an org audit log, per-user usage & cost, skill governance, knowledge sources, and agent access.
📋
Audit & Authority Policy
Every tool call, file write, and approval is logged. Declarative rules decide what runs automatically vs. what needs a human — per action, per relationship.
🔑
Invite-Only Access
No default logins. First registrant becomes admin; everyone else joins by single-use invite code. JWT auth, encrypted credential vault, AES-256 secrets.
🔗
Agent API (A2A)
Expose your agents to other systems over authenticated agent-to-agent calls. Service-account consumers, metered usage, CSV billing, hostile-agent detection.
📚
Org Knowledge Sources
Sync Notion, Slack, and Drive into a shared knowledge base on a schedule. Every member's agent answers grounded in your company's own documents.
Done-For-You Vertical
A real example: a bookkeeper that runs itself.
The same platform, pointed at one job. Email in → reconciled books → P&L emailed back. This is what "AI that works for your clients" looks like in production.
1
Client emails statements
Each client gets their own isolated tenant and a dedicated address. They email bank and credit-card CSVs — attachments are pulled in automatically.
2
The agent reconciles
The bookkeeper profile categorizes transactions, reconciles against prior balances, and flags anomalies — in a sandbox, building a live-formula workbook.
3
You approve
Before anything reaches the client, you get a one-tap human approval. Review the numbers, then release — nothing goes out on its own.
4
Client gets their P&L
A finished Excel workbook with live formulas and a zero-difference reconciliation lands back in the client's inbox. Month-end close, on autopilot.
Early Access
Be First for V2
📅 Private beta — invites rolling out🔒 Not open-source — invitation only
Secure your spot now. Early access members get first invites, exclusive onboarding, and direct founder contact.
🔥 Spots are limited — once beta fills, waitlist closes.
Deployment
One command. Done.
Single Docker Compose. One API key minimum — or use local Ollama for free.
3
generations: agent, workforce, twin
17
MCP servers (V1) → 36 (V2)
16
Agent profiles
5
Messaging platforms
1
Command to deploy
0
Keys needed (Ollama)
Install the open-source edition (V1) on your machine
